One or more aspects of the invention relate, in general, to multiprocessing computing environments, and in particular, to managing transactional and non-transactional store observability.
An enduring challenge in multiprocessor programming is that of updates to the same storage location by multiple central processing units (CPUs). Many instructions that update storage locations, including even simple logical operations, such as AND, do so with multiple accesses to the location. For instance, first, the storage location is fetched, and then, the updated result is stored back.
In order for multiple CPUs to safely update the same storage location, access to the location is serialized. One instruction, the TEST AND SET instruction, introduced with the S/360 architecture formerly offered by International Business Machines Corporation, provided an interlocked update of a storage location. Interlocked update means that, as observed by other CPUs and the input/output (I/O) subsystem (e.g., channel subsystem), the entire storage access of the instruction appears to occur atomically. Later, the S/370 architecture offered by International Business Machines Corporation introduced the COMPARE AND SWAP and COMPARE DOUBLE AND SWAP instructions that provide a more sophisticated means of performing interlocked update, and allow the implementation of what is commonly known as a lock word (or semaphore). Recently added instructions have provided additional interlocked-update capabilities, including COMPARE AND SWAP AND PURGE, and COMPARE AND SWAP AND STORE.
More complex program techniques may require the interlocked update of multiple storage locations, such as when adding an element to a doubly-linked list. In such an operation, both a forward and backward pointer are to appear to be simultaneously updated, as observed by other CPUs and the I/O subsystem. In order to affect such a multiple location update, the program is forced to use a separate, single point of serialization, such as a lock word. However, lock words may provide a much courser level of serialization than is warranted; for example, the lock words may serialize an entire queue of millions of elements, even though only two elements are being updated. The program may structure the data to use finer-grained serialization (e.g., a hierarchy of lock points), but that introduces additional problems, such as potential deadlock situations if the hierarchy is violated, and recovery issues if the program encounters an error while holding one or more locks or if the lock cannot be acquired.
In addition to the above, there are numerous scenarios where a program may execute a sequence of instructions that may or may not result in an exception condition. If no exception condition occurs, then the program continues; however, if an exception is recognized, then the program may take corrective action to eliminate the exception condition. Java®, as one example, can exploit such execution in, for instance, speculative execution, partial in-lining of a function, and/or in the re-sequencing of pointer null checking.
In classic operating system environments, such as z/OS and its predecessors offered by International Business Machines Corporation, the program establishes a recovery environment to intercept any program-exception condition that it may encounter. If the program does not intercept the exception, the operating system typically abnormally terminates the program for exceptions that the operating system is not prepared to handle.